Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Chen, Wenlin, Ge, Hong
Format: Preprint
Veröffentlicht: 2023
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2305.15912
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866912070465748992
author Chen, Wenlin
Ge, Hong
author_facet Chen, Wenlin
Ge, Hong
contents We introduce a novel approach for analyzing the training dynamics of ReLU networks by examining the characteristic activation boundaries of individual ReLU neurons. Our proposed analysis reveals a critical instability in common neural network parameterizations and normalizations during stochastic optimization, which impedes fast convergence and hurts generalization performance. Addressing this, we propose Geometric Parameterization (GmP), a novel neural network parameterization technique that effectively separates the radial and angular components of weights in the hyperspherical coordinate system. We show theoretically that GmP resolves the aforementioned instability issue. We report empirical results on various models and benchmarks to verify GmP's advantages of optimization stability, convergence speed and generalization performance.
format Preprint
id arxiv_https___arxiv_org_abs_2305_15912
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Neural Characteristic Activation Analysis and Geometric Parameterization for ReLU Networks
Chen, Wenlin
Ge, Hong
Machine Learning
We introduce a novel approach for analyzing the training dynamics of ReLU networks by examining the characteristic activation boundaries of individual ReLU neurons. Our proposed analysis reveals a critical instability in common neural network parameterizations and normalizations during stochastic optimization, which impedes fast convergence and hurts generalization performance. Addressing this, we propose Geometric Parameterization (GmP), a novel neural network parameterization technique that effectively separates the radial and angular components of weights in the hyperspherical coordinate system. We show theoretically that GmP resolves the aforementioned instability issue. We report empirical results on various models and benchmarks to verify GmP's advantages of optimization stability, convergence speed and generalization performance.
title Neural Characteristic Activation Analysis and Geometric Parameterization for ReLU Networks
topic Machine Learning
url https://arxiv.org/abs/2305.15912